Porównanie cząstkowych modeli parametrycznych w zadaniu detekcji uszkodzeń sieci gazowej

Zreferowano badania detekcji uszkodzeń gazociągu z użyciem cząstkowych modeli parametrycznych. Stosując trzy metody modelowania: addytywne modele regresyjne (najnowszą z badanych technik), sztuczne sieci neuronowe oraz układy rozmyte typu TSK opracowano aproksymacje ciśnień w węzłach sieci. Modele testowano w zadaniu detekcji wycieku oraz uszkodzenia czujnika pomiarowego. Wszystkie modele zapewniały dużą dokładność aproksymacji ciśnienia w poprawnych stanach pracy, wykazując także bardzo skuteczną detekcję uszkodzeń czujników pomiarowych ciśnień, natomiast w sytuacji symulowanych wycieków ich przydatność w detekcji była znacznie mniejsza.

The results of faults detection [1, 2, 3, 4, 5] in a gas system network (Fig. 1) with use of parametric partial models [6, 7, 8] are presented in the paper. This is a new approach to the task with use of exploratory data analysis [10, 11, 17] and partial models. Three techniques were used to build models of pressure in network nodes: additive regression (ADD - new method of modelling [10, 11, 12, 13, 14, 15]), artificial neural networks (ANN) [16, 17, 18] and TSK fuzzy logic modelling [8, 16, 17]. The measured pressures in adjacent nodes as well cumulative flows in the main line (from global analytical model [9]) of gasoline were the inputs of the models. For the analysed stations (in parts A and B marked in Fig. 1) a set of test failures in the form of leaks and damage of pressure sensors is given in Tab. 1.Using trial and error method, by evaluating the effectiveness of fault detection, there were obtained structures of models of different complexity for individual modelling techniques: ADD - presented by equations (1) and (2), ANN- (3) and (4), TSK- (5) and (6). The model order is not greater than 2. The exemplary results of leak detection with use of particular models are shown in Figs. 3, 5, 7 and of sensor fault detection in Figs. 4, 6, 8. In the conclusions there is summarised the relative accuracy of models (in Table 2), the relative normalized values of the studied residues of leaks - Tab.3 and the pressure sensor failures - Tab. 4. All models provided highly precise pressure approximation in non-fault states, but TSK and ADD models turned out to be the more accurate. Additionally, all of them were effective in case of pressure sensor fault detection, however, in case of simulated leakages their usefulness was much lower.